I'm an absolutely newbie to signal processing. I'm trying to classify EMG signals which are very noisy (decibel values are more than -70 dB in some cases). After applying EMD technique these values are improved to -30 dB to -40 dB.

My question is :

I want to classify these EMG signals. It's a binary classification problem. If these noisy signals are fed to a complex classification algorithm like CNN, it will learn the features from the noisy signals only. This knowledge it will apply to classify unknown noisy signals. As all the samples in my dataset are affected by the same kinds of noises, do the noises affect too much the final validation accuracy ? Or, should I focus in denoising further?


1 Answer 1


-30 dB is still very noisy.

If you've had success with EMD, I'd try an inspired transform that's improved on it: synchrosqueezing. Whether it's best to denoise before classifying depends on amount of available data: most denoising will throw away some valuable information, but also make the task easier for a classifier. If there's lots of data, don't denoise.

I'd also try scattering which'll yield timeshift-invariant features (lecture); with this much noise one can't hope for precise temporal localization anyway.

Finally I'd look for architectures pretrained on a similar noise profile (or sufficiently powerful to generalize) and attempt transfer learning. But the most promising approach would be data-oriented: find less noisy samples.

  • 1
    $\begingroup$ Thanks for the pointers. I'll check them I'm thinking to test these samples on some ML & DL models. After seeing the results, will decide on further denoising. $\endgroup$
    – Debbie
    Jul 22, 2021 at 9:59

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.